Concept Drift Detection on Streaming Data with Dynamic Outlier Aggregation

Many processes no matter what kind are regularly changing over time, adapting themselves to external and internal circumstances. Analyzing them in a streaming context is a very demanding task. Particularly the detection and classification of significant deviations is important to be able to re-integrate these possible micro-processes. Assuming a deviation of a certain process, the significance is implicitly given when a high number of instances contain this deviation similarly. To enhance a process the integration of or preventive measures against those anomalies is of high interest for all stakeholders as the actual process core gets discovered more and more in detail. Considering various areas of application, we focus on previously neglected but potentially significant anomalies like small changes in the disease process of a virus infection that has to be discovered to develop an appropriate reaction mechanism. We concentrate on non-conforming traces of a stream on which we compute a local outlier factor. This allows us to detect relations between traces based on changing outlier scores. Hence, hereby connected traces are clusters with which we achieve the detection of concept drift. We evaluate our approach on a synthetic event log and a real-world dataset corresponding to a process representing building permit applications which emphasizes the extensive applicability.